Inspired by Developmental Psychology, Deepmind Researchers Develop a Deep Learning Model to Learn Intuitive Physics

Key Contributions: (1) Researchers introduce and open-source a machine-learning dataset designed to evaluate the conceptual understanding of intuitive physics, adopting the violation-of-expectation (VoE) paradigm from developmental psychology, (2) They build a deep-learning system that learns intuitive physics directly from visual data

Atari video games, board games like chess and go, scientific difficulties like protein folding, and language modeling—all of these activities are now being mastered by artificial intelligence (AI), which has made astounding strides in recent years. Nevertheless, it has become increasingly obvious that something essential is still lacking as a result of success in these specific fields. Modern AI systems are particularly challenged by the “common sense” information that directs prediction, inference, and behavior in commonplace human contexts.

In a recent study, Deepmind researchers concentrated on one specific area of common-sense knowledge: intuitive physics, the conceptual network that supports understanding of the characteristics and interactions of macroscopic objects. It goes without saying that intuitive physics is crucial to every practical activity, but it also serves as a foundation for conceptual knowledge and compositional representation in general. However, despite great effort, recent developments in AI have yet to produce a system that has a comprehension of intuitive physics equivalent to even very young children.

The team was frequently inspired by developmental psychology, where the acquisition of intuitive physics knowledge has been a major area of study in order to explore richer common sense physical intuition in AI systems. The basic discovery of the developmental literature—that physics is comprehended at the level of discrete objects and their interactions—was integrated into a deep-learning system by researchers.

Additionally, researchers used developmental psychology to address the issue of behaviorally determining whether an AI system is knowledgeable about intuitive physics.

Developmental psychologists have used two guiding concepts to create behavioral probes for studies on children. First, the foundation of intuitive physics is a collection of distinct ideas that can be distinguished, operationalized, and individually investigated. The work is quite different from conventional approaches in AI for learning intuitive physics because it specifically targets discrete concepts. These approaches typically measure progress through video or state prediction metrics, binary outcome prediction, question-answering performance, or high reward in reinforcement learning tasks.

The second rule that developmental psychologists employ when examining physical concepts is that having a physical concept is equivalent to developing a set of expectations about how the future might play out. The violation-of-expectation (VoE) paradigm, a way for gauging understanding of a particular physical notion, arises from this conceptual scaffolding.

Researchers offer visually identical arrays that are either consistent or inconsistent with that physical idea to infants using the VoE paradigm to probe for a particular concept. If infants are more surprised by the impossible array, this indicates that their expectations, which were based on their understanding of the physical notion under investigation, were not met.

Researchers evaluated the model using an externally defined dataset with unique appearances, shapes, and dynamics without any retraining to test the network’s generalization ability. The findings revolve around four main findings. First, despite having been trained on video data in which the precise probe events did not occur, the object-based model demonstrated robust VoE effects across all five concepts examined. Second, the researchers discovered that the VoE effects seen in the model lessened or vanished in well-matched models that did not use object-centered representations, in line with expectations based on the developmental literature. Third, researchers found that even with just 28 hours of visual training data, it was possible to get strong VoE effects. The model generalized, according to the researchers, to a dataset that was separately created and contained fresh item forms and dynamics.


Numerous forms of “common sense” reasoning have become a significant research focus for AI. A recent study by Deepmind researchers detailed advancements made in the development of deep learning systems that create knowledge in the area of intuitive physics, a particular but fundamental area of common sense, based on perceptual experience. To design a model capable of learning intuitive physics, researchers borrowed inspiration from developmental psychology to monitor progress quantitatively and systematically. The researchers added object-centric representation and computation to the model, which was directly influenced by descriptions of early intuitive physics.

The research significantly builds upon and is closely related to a number of earlier works in computational cognitive science and artificial intelligence. The team believes it would be exciting to expand on the current modeling work in order to engage with important developmental psychology issues even more directly.

This Article is written as a summary article by Marktechpost Staff based on the research paper 'Intuitive physics learning in a deep-learning model inspired by developmental psychology'. All Credit For This Research Goes To Researchers on This Project. Checkout the paper.

Please Don't Forget To Join Our ML Subreddit